Continuous Field Reconstruction from Sparse Observations with Implicit Neural Networks

Authors: Xihaier Luo, Wei Xu, Balu Nadiga, Yihui Ren, Shinjae Yoo

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In experimental evaluations, the proposed model outperforms recent INR methods, offering superior reconstruction quality on simulation data from a stateof-the-art climate model and a second dataset that comprises ultra-high resolution satellite-based sea surface temperature fields.
Researcher Affiliation Academia Xihaier Luo, Wei Xu, Yihui Ren, Shinjae Yoo Brookhaven National Laboratory {xluo,xuw,yren,sjyoo}@bnl.gov Balasubramanya Nadiga Los Alamos National Laboratory {balu}@lanl.gov
Pseudocode No No structured pseudocode or algorithm blocks were found in the paper.
Open Source Code Yes [Project Website: data & code]
Open Datasets Yes For these experiments, we evaluate the model s performance on two challenging datasets (additionally detailed in Appendix A): Simulation-based data. The Community Earth System Model version 2 (CESM2) (Danabasoglu et al., 2020)... The U.S. Department of Energy s Program for Climate Model Diagnosis and Intercomparison (PCMDI) provides coordinating support and development of software infrastructure in partnership with the Global Organization for Earth System Science Portals to provide the data at pcmdi.llnl.gov. Satellite-based data. Sea surface temperature data are derived from both a retrospective dataset with a four-day latency and a near-real-time dataset with a one-day latency (Martin et al., 2012)... NASA provides full and open access to this data under its Earth Science Data Systems (ESDS) Program at earthdata.gov.
Dataset Splits No No explicit percentage or sample count was provided for a dedicated validation dataset split. The paper mentions 'testing and validation' in Appendix A but only specifies 'training procedure involves using partial observations sampled from the complete state... Testing employs the complete state (s = 100%).' for split details.
Hardware Specification Yes We conduct all experiments using Py Torch Lightning, repeating each experiment 10 times on a single NVIDIA A100 40 GB GPU.
Software Dependencies No The paper mentions 'PyTorch Lightning' but does not provide specific version numbers for any software dependencies.
Experiment Setup Yes All models are trained with an L2 loss function for 200 epochs, employing the Adam W optimizer (Loshchilov & Hutter, 2019) with an initial learning rate of 0.001. We apply a learning rate decay of 0.99 for each parameter group of every epoch. Due to variations in resolution between earth simulation and satellite imagery data, we set batch sizes to 16 and 2, respectively.